Psychology Seminar Psych 406 Dr. Jeffrey Leitzel

Transcription

1 Psychology Seminar Psych 406 Dr. Jeffrey Leitzel Structural Equation Modeling Topic 1: Correlation / Linear Regression Outline/Overview Correlations (r, pr, sr) Linear regression Multiple regression interpreting b coefficients ANOVA model test R 2 Diagnostics Distance(Discrepancy), Leverage, Influence, & Multicollinearity The correlation coefficient (Pearson r) Continuous IV & DV (Interval/Ratio level data) Dichotomous variables Strength of linear relationship between variables r 2 coefficient of determination r Z xz y N 1 1

2 Partial and semipartial correlations Partial Correlation r between two variables when one or more other variables is partialled out of both X and Y example: experimenter interested in investigating the relationship between income and success in college measured both variables, ran correlation it was statistically significant he began repeating these results to all of his students do well in college = large salaries what might the bright student in the back of one of his classes have raised as an issue? Partial correlation (cont.) IQ as a third variable effecting both college success and salary earned The partial correlation between college success and salary with IQ partialled out of both variables salary(y) IQ(x 2 ) salary(y) IQ(x 2 ) gpa(x 1 ) gpa(x 1 ) Partial correlation (cont.) correlation between the residuals is the partial correlation between income and success, partialling out IQ generally represented by r y1.23 p subscripts to the left of the dot represent variables being correlated those to the right of the dot are those being partialled out r y1.2 =correlation IQ(x 2 ) between salary and salary(y) GPA with IQ partialled D A out pr 2 B =c/(c + d) gpa(x 1 ) C 2

3 Semipartial correlation also called the part correlation, far more routinely used r y(1.2) represents the correlation between the criterion (y) and a partialled predictor variable the partial correlation has variable x 2 partialled out of both y and x 1 for the semipartial, x 2 is partialled only out of x 1 the correlation between y and the residuals from x 1 predicted by x 2 correlation between y and the part of x 1 that is independent of x 2 r y(1.2) =correlation between salary and IQ(x 2 ) GPA with IQ partialled out salary(y) sr 2 =c/(a + b + c + d) D A B gpa(x 1 ) C Semipartial correlation (cont.) can rearrange semipartial correlation formula: R 2 y.12 = r 2 y2 + r 2 y(1.2) R (Multiple correlation) based on information from variable 1 plus additional nonredundant information from variable 2 through variable p so: R 2 y.123 p = r 2 y1 + r 2 y(2.1) + r 2 y(3.12) + + r 2 y(p.123 p-1) If the predictors were completely independent of one another, there would be no shared variance to partial out would simply sum the r 2 for each X variable with Y to get the Multiple R 2 Simple linear regression Finding the equation for the line of best fit Least squares property Minimizes errors of estimation The line will have the form: y = a + bx Where: y' = predicted value of y a = intercept of the line b = slope of the line x = score of x we are using to predict y Multiple regression an extension of this 3

4 Multiple regression models Simple linear regression models generally represent an oversimplification of reality Usually unreasonable to think in terms of a single cause for any variable of interest Our equation for simple linear regression: y = a + bx becomes: y = a+b 1 x 1 +b 2 x 2 + +b k x k Multiple regression (cont.) y =a+b 1 x 1 +b 2 x 2 + +b k x k Where y = dependent variable (what we are predicting) x 1, x 2,, x k = predictors (independent variables or regressors) b 1, b 2,, b k = regression coefficients associated with the k predictors a = y intercept The predictors (x s) may represent interaction terms, cross products, powers, logs or other functions that provide predictive power Interpreting b coefficients If x 1 is held constant, every unit change in x 2 results in a b 2 unit change in y The constant, a (y intercept) does not necessarily have a meaningful interpretation any time 0 is outside the reasonable range for predictors the intercept will be essentially meaningless 4

5 Overall ANOVA Testing overall model adequacy Null hypothesis H 0 : b 1 =b 2 = =b k =0 Alternative hypoth H 1 : any b <> 0 F statistic (ANOVA) = MS regression /MS error MS regression = SS regression / k MS error = SS error / n-(k+1) Where n = number of observations k = number of parameters in model (excluding a) ANOVA (cont.) SS regression = (y - M y ) 2 SS error = (y - y ) 2 When the overall regression model provides no predictive power the two mean square quantities will be about equal We are testing an F statistic with (k, n-(k+1)) degrees of freedom ANOVA in multiple regression 5

6 R 2 and Adjusted R 2 Multiple R 2, coefficient of determination Total proportion of variance accounted for by all predictors in the model High R 2, values do not necessarily mean that a model will be useful for predicting Y Overfitting (adding essentially irrelevant predictors) can result in a high R 2 in the absence of any predictive power If we fit a regression model with n-1 predictors R will always be = 1.0 unless we have a rare data set where two cases are identical on all predictors but differ on Y value Adjusted R 2 takes into account our n and the number of predictors we have used in the model Misuses/problems Multicollinearity exists when two or more independent variables (predictors) contribute redundant information to the model can think of conceptually as a problem in assigning unique variance components to variable creates computational problems and coefficients that do not make sense resolve by removing redundant predictor(s) Predicting outside the range of data that was used to generate the regression model positive linear relationship relationship may change-resulting in substantial error of estimation Failure to explore alternative models need to be familiar with data examine relationships between variables be aware of potential interactions and include in model if necessary try different models Standardized betas with more than one independent variable: b coefficients cannot be directly compared due to different scales across measures standardized Beta coefficients allow us to compare the relative predictive power of variables in the equation we get standardized Betas by converting all of our predictors to z-scores and running the regression 6

7 Regression diagnostics help us assess the validity of our conclusions and their accuracy careful screening of the data cannot be overemphasized Outliers outliers: data points that lie outside the general linear pattern (midline is the regression line) removal of outliers can dramatically affect the performance of a regression model should be removed if there is reason to believe that other variables not in the model explain why the cases are unusual these cases may need a separate model Outliers (cont.) outliers may also suggest that additional explanatory variables need to be brought into the model multivariate outliers can be difficult if not impossible to identify by visual inspection ex. 115 lbs. and 6 feet tall either observation alone is not unusual, but together they are. unusual combinations might include odd pairings of scores on separate measures of ability or different indices of production multivariate outliers represent results that when considered as a whole do not make sense Distance, leverage, & influence most common measure of distance (also called discrepancy) is the simple residual distance between any point and the regression surface identifies outliers on the dependent (y) variable leverage statistic, h, also called hat-value, identifies cases which may influence the regression model more than others with reasonably large n ranges from essentially 0 (no influence on the model, actually the min is 1/n) to 1 7

8 Distance, leverage, & influence (cont.) leverage identifies outliers in the dv s <.2 fine, >.5 case should be examined cases that are high on distance or leverage can strongly influence the regression but do not necessarily do so points that are relatively high on both distance and leverage are very likely to have strong influence Influence on regression line Influence on regression line 8

9 Influence on regression line Cook s distance (D) most common measure of influence of a case it is a function of the squared change that would occur in y if the observation were removed from the data interested in finding cases with D values much larger than the rest cut-off influential cases, D greater than 4/(n - k - 1) where n is the number of cases and k is the number of independents Cook's distance obtained by issuing postestimation command in Stata Regression diagnostics 9

10 Diagnostics put to use Multicollinearity intercorrelation of independent variables R 2 s near 1 violate assumption of no perfect collinearity high R 2 s increase the standard error of the beta coefficients and make assessment of the unique role of each independent difficult or impossible simple correlations tell something about multicollinearity preferred method of assessing multicollinearity is to regress each independent on all the other independent variables in the equation Multicollinearity (cont.) inspection of the correlation matrix reveals only bivariate multicollinearity, for bivariate correlations > 0.90 to assess multivariate multicollinearity, one uses tolerance or VIF may not be any extremely high bivariate correlations if any variable can be represented as a linear combination of other variables in the model, perfect multicollinearity exists 10

11 Tolerance/Variance Inflation Factor (VIF) Tolerance 1 - R 2 for the regression of that independent variable on all other independents, ignoring the dependent as many tolerance coefficients as there are independents higher intercorrelation of the independents, the tolerance will approach zero part of the denominator for calculating the confidence limits on the b (partial regression) coefficient Variance inflation factor (VIF) reciprocal of tolerance when VIF is high there is high multicollinearity and instability of the b coefficients Variance inflation factor (VIF) Rj Tolerance VIF Impact on SEb